Solved CPU usage on idle

[drohaw]

I have two small VPS instances, one running FreeBSD and the other running Linux. The FreeBSD sits at 2% CPU usage when the system is doing literally nothing, whereas Linux instance is 0%.

Any suggestions for system tuning to get this down to 0%? From top output, the process using the most CPU time is rand harvest -- is that just churning even though the system is idle? If so, anything I can do about it? I know 2% is small, but I'd like to understand what the system is doing, and why Linux doesn't have this effect.

 

[diizzy]

Might be related to https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=173541

 

[SirDice]

What version of FreeBSD?

 

[a6h]

ZFS? If so, post the output of the zpool status. It may have something to do with zpool-resilver(8)

 

[PMc]

My machine does never do really "nothing", but I observed this:

By default my kernels get compiled with HZ=200 (aka sysctl kern.hz). I changed a bhyve instance to kern.hz=4000, and consequentially the idle cpu (i5-3570T) power usage changed from 19W to 26W (as shown with sysutils/intel-pcm).

Logically follows from this that even with low HZ there will be some CPU usage (maybe 2% ?).

Background: what I am focused on is dummynet(4). And dummynet is also the primary consumer of the power. To work properly, dummynet seems to need a very high HZ value. I do not know how things would be with dummynet unused (i.e. no ipfw firewall) or not compiled in.

 

[diizzy]

Most platforms use 1000 as default which should be enough for most usecases including altq/dummynet

 

[drohaw]

Ah, that explains it. So Linux has a "CONFIG_NO_HZ_IDLE" feature, which puts the kernel into tickless mode when the system is idle. FreeBSD doesn't have that, so it's using 2% of the CPU waking itself up 1000 times per second for no reason. Anyway, thanks for the help, and marking this solved.

 

[Alain De Vos]

There is a reason why kern.hz is 1000.
Why is 10 not a good value for number of clock ticks per second ?
Or why is 100000 not a good value for number of clock ticks per second ?

 

[drohaw]

There is a reason why kern.hz is 1000.
Why is 10 not a good value for number of clock ticks per second ?
Or why is 100000 not a good value for number of clock ticks per second ?

Sure -- I didn't mean to imply that there is anything wrong with the FreeBSD implementation. Just trying to figure out discrepancy between my Linux and FreeBSD instances.

 

[trev]

My VPS instances at Binary Lane sit at between 99.2% and 100% idle. My VPS instance at Vultr sits between 98.1% and 100% idle.

I no longer recall which one, but one of them was always at 5% until I lodged a support ticket and they finagled something.